4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/module.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <asm/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Workqueue for cpuset related tasks.
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
69 static struct workqueue_struct *cpuset_wq;
72 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
76 int number_of_cpusets __read_mostly;
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys;
82 /* See "Frequency meter" comments, below. */
85 int cnt; /* unprocessed events count */
86 int val; /* most recent output value */
87 time_t time; /* clock (secs) when val computed */
88 spinlock_t lock; /* guards read or write of above */
92 struct cgroup_subsys_state css;
94 unsigned long flags; /* "unsigned long" so bitops work */
95 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
98 struct cpuset *parent; /* my parent */
100 struct fmeter fmeter; /* memory_pressure filter */
102 /* partition number for rebuild_sched_domains() */
105 /* for custom sched domain */
106 int relax_domain_level;
108 /* used for walking a cpuset hierarchy */
109 struct list_head stack_list;
112 /* Retrieve the cpuset for a cgroup */
113 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
115 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset *task_cs(struct task_struct *task)
122 return container_of(task_subsys_state(task, cpuset_subsys_id),
126 /* bits in struct cpuset flags field */
132 CS_SCHED_LOAD_BALANCE,
137 /* convenient tests for these bits */
138 static inline int is_cpu_exclusive(const struct cpuset *cs)
140 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
143 static inline int is_mem_exclusive(const struct cpuset *cs)
145 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
148 static inline int is_mem_hardwall(const struct cpuset *cs)
150 return test_bit(CS_MEM_HARDWALL, &cs->flags);
153 static inline int is_sched_load_balance(const struct cpuset *cs)
155 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
158 static inline int is_memory_migrate(const struct cpuset *cs)
160 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
163 static inline int is_spread_page(const struct cpuset *cs)
165 return test_bit(CS_SPREAD_PAGE, &cs->flags);
168 static inline int is_spread_slab(const struct cpuset *cs)
170 return test_bit(CS_SPREAD_SLAB, &cs->flags);
173 static struct cpuset top_cpuset = {
174 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
178 * There are two global mutexes guarding cpuset structures. The first
179 * is the main control groups cgroup_mutex, accessed via
180 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
181 * callback_mutex, below. They can nest. It is ok to first take
182 * cgroup_mutex, then nest callback_mutex. We also require taking
183 * task_lock() when dereferencing a task's cpuset pointer. See "The
184 * task_lock() exception", at the end of this comment.
186 * A task must hold both mutexes to modify cpusets. If a task
187 * holds cgroup_mutex, then it blocks others wanting that mutex,
188 * ensuring that it is the only task able to also acquire callback_mutex
189 * and be able to modify cpusets. It can perform various checks on
190 * the cpuset structure first, knowing nothing will change. It can
191 * also allocate memory while just holding cgroup_mutex. While it is
192 * performing these checks, various callback routines can briefly
193 * acquire callback_mutex to query cpusets. Once it is ready to make
194 * the changes, it takes callback_mutex, blocking everyone else.
196 * Calls to the kernel memory allocator can not be made while holding
197 * callback_mutex, as that would risk double tripping on callback_mutex
198 * from one of the callbacks into the cpuset code from within
201 * If a task is only holding callback_mutex, then it has read-only
204 * Now, the task_struct fields mems_allowed and mempolicy may be changed
205 * by other task, we use alloc_lock in the task_struct fields to protect
208 * The cpuset_common_file_read() handlers only hold callback_mutex across
209 * small pieces of code, such as when reading out possibly multi-word
210 * cpumasks and nodemasks.
212 * Accessing a task's cpuset should be done in accordance with the
213 * guidelines for accessing subsystem state in kernel/cgroup.c
216 static DEFINE_MUTEX(callback_mutex);
219 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
220 * buffers. They are statically allocated to prevent using excess stack
221 * when calling cpuset_print_task_mems_allowed().
223 #define CPUSET_NAME_LEN (128)
224 #define CPUSET_NODELIST_LEN (256)
225 static char cpuset_name[CPUSET_NAME_LEN];
226 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
227 static DEFINE_SPINLOCK(cpuset_buffer_lock);
230 * This is ugly, but preserves the userspace API for existing cpuset
231 * users. If someone tries to mount the "cpuset" filesystem, we
232 * silently switch it to mount "cgroup" instead
234 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
235 int flags, const char *unused_dev_name, void *data)
237 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
238 struct dentry *ret = ERR_PTR(-ENODEV);
242 "release_agent=/sbin/cpuset_release_agent";
243 ret = cgroup_fs->mount(cgroup_fs, flags,
244 unused_dev_name, mountopts);
245 put_filesystem(cgroup_fs);
250 static struct file_system_type cpuset_fs_type = {
252 .mount = cpuset_mount,
256 * Return in pmask the portion of a cpusets's cpus_allowed that
257 * are online. If none are online, walk up the cpuset hierarchy
258 * until we find one that does have some online cpus. If we get
259 * all the way to the top and still haven't found any online cpus,
260 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
261 * task, return cpu_online_map.
263 * One way or another, we guarantee to return some non-empty subset
266 * Call with callback_mutex held.
269 static void guarantee_online_cpus(const struct cpuset *cs,
270 struct cpumask *pmask)
272 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
275 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
277 cpumask_copy(pmask, cpu_online_mask);
278 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
282 * Return in *pmask the portion of a cpusets's mems_allowed that
283 * are online, with memory. If none are online with memory, walk
284 * up the cpuset hierarchy until we find one that does have some
285 * online mems. If we get all the way to the top and still haven't
286 * found any online mems, return node_states[N_HIGH_MEMORY].
288 * One way or another, we guarantee to return some non-empty subset
289 * of node_states[N_HIGH_MEMORY].
291 * Call with callback_mutex held.
294 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
296 while (cs && !nodes_intersects(cs->mems_allowed,
297 node_states[N_HIGH_MEMORY]))
300 nodes_and(*pmask, cs->mems_allowed,
301 node_states[N_HIGH_MEMORY]);
303 *pmask = node_states[N_HIGH_MEMORY];
304 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
308 * update task's spread flag if cpuset's page/slab spread flag is set
310 * Called with callback_mutex/cgroup_mutex held
312 static void cpuset_update_task_spread_flag(struct cpuset *cs,
313 struct task_struct *tsk)
315 if (is_spread_page(cs))
316 tsk->flags |= PF_SPREAD_PAGE;
318 tsk->flags &= ~PF_SPREAD_PAGE;
319 if (is_spread_slab(cs))
320 tsk->flags |= PF_SPREAD_SLAB;
322 tsk->flags &= ~PF_SPREAD_SLAB;
326 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
328 * One cpuset is a subset of another if all its allowed CPUs and
329 * Memory Nodes are a subset of the other, and its exclusive flags
330 * are only set if the other's are set. Call holding cgroup_mutex.
333 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
335 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
336 nodes_subset(p->mems_allowed, q->mems_allowed) &&
337 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
338 is_mem_exclusive(p) <= is_mem_exclusive(q);
342 * alloc_trial_cpuset - allocate a trial cpuset
343 * @cs: the cpuset that the trial cpuset duplicates
345 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
347 struct cpuset *trial;
349 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
353 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
357 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
363 * free_trial_cpuset - free the trial cpuset
364 * @trial: the trial cpuset to be freed
366 static void free_trial_cpuset(struct cpuset *trial)
368 free_cpumask_var(trial->cpus_allowed);
373 * validate_change() - Used to validate that any proposed cpuset change
374 * follows the structural rules for cpusets.
376 * If we replaced the flag and mask values of the current cpuset
377 * (cur) with those values in the trial cpuset (trial), would
378 * our various subset and exclusive rules still be valid? Presumes
381 * 'cur' is the address of an actual, in-use cpuset. Operations
382 * such as list traversal that depend on the actual address of the
383 * cpuset in the list must use cur below, not trial.
385 * 'trial' is the address of bulk structure copy of cur, with
386 * perhaps one or more of the fields cpus_allowed, mems_allowed,
387 * or flags changed to new, trial values.
389 * Return 0 if valid, -errno if not.
392 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
395 struct cpuset *c, *par;
397 /* Each of our child cpusets must be a subset of us */
398 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
399 if (!is_cpuset_subset(cgroup_cs(cont), trial))
403 /* Remaining checks don't apply to root cpuset */
404 if (cur == &top_cpuset)
409 /* We must be a subset of our parent cpuset */
410 if (!is_cpuset_subset(trial, par))
414 * If either I or some sibling (!= me) is exclusive, we can't
417 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
419 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
421 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
423 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
425 nodes_intersects(trial->mems_allowed, c->mems_allowed))
429 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
430 if (cgroup_task_count(cur->css.cgroup)) {
431 if (cpumask_empty(trial->cpus_allowed) ||
432 nodes_empty(trial->mems_allowed)) {
442 * Helper routine for generate_sched_domains().
443 * Do cpusets a, b have overlapping cpus_allowed masks?
445 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
447 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
451 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
453 if (dattr->relax_domain_level < c->relax_domain_level)
454 dattr->relax_domain_level = c->relax_domain_level;
459 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
463 list_add(&c->stack_list, &q);
464 while (!list_empty(&q)) {
467 struct cpuset *child;
469 cp = list_first_entry(&q, struct cpuset, stack_list);
472 if (cpumask_empty(cp->cpus_allowed))
475 if (is_sched_load_balance(cp))
476 update_domain_attr(dattr, cp);
478 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
479 child = cgroup_cs(cont);
480 list_add_tail(&child->stack_list, &q);
486 * generate_sched_domains()
488 * This function builds a partial partition of the systems CPUs
489 * A 'partial partition' is a set of non-overlapping subsets whose
490 * union is a subset of that set.
491 * The output of this function needs to be passed to kernel/sched.c
492 * partition_sched_domains() routine, which will rebuild the scheduler's
493 * load balancing domains (sched domains) as specified by that partial
496 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
497 * for a background explanation of this.
499 * Does not return errors, on the theory that the callers of this
500 * routine would rather not worry about failures to rebuild sched
501 * domains when operating in the severe memory shortage situations
502 * that could cause allocation failures below.
504 * Must be called with cgroup_lock held.
506 * The three key local variables below are:
507 * q - a linked-list queue of cpuset pointers, used to implement a
508 * top-down scan of all cpusets. This scan loads a pointer
509 * to each cpuset marked is_sched_load_balance into the
510 * array 'csa'. For our purposes, rebuilding the schedulers
511 * sched domains, we can ignore !is_sched_load_balance cpusets.
512 * csa - (for CpuSet Array) Array of pointers to all the cpusets
513 * that need to be load balanced, for convenient iterative
514 * access by the subsequent code that finds the best partition,
515 * i.e the set of domains (subsets) of CPUs such that the
516 * cpus_allowed of every cpuset marked is_sched_load_balance
517 * is a subset of one of these domains, while there are as
518 * many such domains as possible, each as small as possible.
519 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
520 * the kernel/sched.c routine partition_sched_domains() in a
521 * convenient format, that can be easily compared to the prior
522 * value to determine what partition elements (sched domains)
523 * were changed (added or removed.)
525 * Finding the best partition (set of domains):
526 * The triple nested loops below over i, j, k scan over the
527 * load balanced cpusets (using the array of cpuset pointers in
528 * csa[]) looking for pairs of cpusets that have overlapping
529 * cpus_allowed, but which don't have the same 'pn' partition
530 * number and gives them in the same partition number. It keeps
531 * looping on the 'restart' label until it can no longer find
534 * The union of the cpus_allowed masks from the set of
535 * all cpusets having the same 'pn' value then form the one
536 * element of the partition (one sched domain) to be passed to
537 * partition_sched_domains().
539 static int generate_sched_domains(cpumask_var_t **domains,
540 struct sched_domain_attr **attributes)
542 LIST_HEAD(q); /* queue of cpusets to be scanned */
543 struct cpuset *cp; /* scans q */
544 struct cpuset **csa; /* array of all cpuset ptrs */
545 int csn; /* how many cpuset ptrs in csa so far */
546 int i, j, k; /* indices for partition finding loops */
547 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
548 struct sched_domain_attr *dattr; /* attributes for custom domains */
549 int ndoms = 0; /* number of sched domains in result */
550 int nslot; /* next empty doms[] struct cpumask slot */
556 /* Special case for the 99% of systems with one, full, sched domain */
557 if (is_sched_load_balance(&top_cpuset)) {
559 doms = alloc_sched_domains(ndoms);
563 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
565 *dattr = SD_ATTR_INIT;
566 update_domain_attr_tree(dattr, &top_cpuset);
568 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
573 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
578 list_add(&top_cpuset.stack_list, &q);
579 while (!list_empty(&q)) {
581 struct cpuset *child; /* scans child cpusets of cp */
583 cp = list_first_entry(&q, struct cpuset, stack_list);
586 if (cpumask_empty(cp->cpus_allowed))
590 * All child cpusets contain a subset of the parent's cpus, so
591 * just skip them, and then we call update_domain_attr_tree()
592 * to calc relax_domain_level of the corresponding sched
595 if (is_sched_load_balance(cp)) {
600 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
601 child = cgroup_cs(cont);
602 list_add_tail(&child->stack_list, &q);
606 for (i = 0; i < csn; i++)
611 /* Find the best partition (set of sched domains) */
612 for (i = 0; i < csn; i++) {
613 struct cpuset *a = csa[i];
616 for (j = 0; j < csn; j++) {
617 struct cpuset *b = csa[j];
620 if (apn != bpn && cpusets_overlap(a, b)) {
621 for (k = 0; k < csn; k++) {
622 struct cpuset *c = csa[k];
627 ndoms--; /* one less element */
634 * Now we know how many domains to create.
635 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
637 doms = alloc_sched_domains(ndoms);
642 * The rest of the code, including the scheduler, can deal with
643 * dattr==NULL case. No need to abort if alloc fails.
645 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
647 for (nslot = 0, i = 0; i < csn; i++) {
648 struct cpuset *a = csa[i];
653 /* Skip completed partitions */
659 if (nslot == ndoms) {
660 static int warnings = 10;
663 "rebuild_sched_domains confused:"
664 " nslot %d, ndoms %d, csn %d, i %d,"
666 nslot, ndoms, csn, i, apn);
674 *(dattr + nslot) = SD_ATTR_INIT;
675 for (j = i; j < csn; j++) {
676 struct cpuset *b = csa[j];
679 cpumask_or(dp, dp, b->cpus_allowed);
681 update_domain_attr_tree(dattr + nslot, b);
683 /* Done with this partition */
689 BUG_ON(nslot != ndoms);
695 * Fallback to the default domain if kmalloc() failed.
696 * See comments in partition_sched_domains().
707 * Rebuild scheduler domains.
709 * Call with neither cgroup_mutex held nor within get_online_cpus().
710 * Takes both cgroup_mutex and get_online_cpus().
712 * Cannot be directly called from cpuset code handling changes
713 * to the cpuset pseudo-filesystem, because it cannot be called
714 * from code that already holds cgroup_mutex.
716 static void do_rebuild_sched_domains(struct work_struct *unused)
718 struct sched_domain_attr *attr;
724 /* Generate domain masks and attrs */
726 ndoms = generate_sched_domains(&doms, &attr);
729 /* Have scheduler rebuild the domains */
730 partition_sched_domains(ndoms, doms, attr);
734 #else /* !CONFIG_SMP */
735 static void do_rebuild_sched_domains(struct work_struct *unused)
739 static int generate_sched_domains(cpumask_var_t **domains,
740 struct sched_domain_attr **attributes)
745 #endif /* CONFIG_SMP */
747 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
750 * Rebuild scheduler domains, asynchronously via workqueue.
752 * If the flag 'sched_load_balance' of any cpuset with non-empty
753 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
754 * which has that flag enabled, or if any cpuset with a non-empty
755 * 'cpus' is removed, then call this routine to rebuild the
756 * scheduler's dynamic sched domains.
758 * The rebuild_sched_domains() and partition_sched_domains()
759 * routines must nest cgroup_lock() inside get_online_cpus(),
760 * but such cpuset changes as these must nest that locking the
761 * other way, holding cgroup_lock() for much of the code.
763 * So in order to avoid an ABBA deadlock, the cpuset code handling
764 * these user changes delegates the actual sched domain rebuilding
765 * to a separate workqueue thread, which ends up processing the
766 * above do_rebuild_sched_domains() function.
768 static void async_rebuild_sched_domains(void)
770 queue_work(cpuset_wq, &rebuild_sched_domains_work);
774 * Accomplishes the same scheduler domain rebuild as the above
775 * async_rebuild_sched_domains(), however it directly calls the
776 * rebuild routine synchronously rather than calling it via an
777 * asynchronous work thread.
779 * This can only be called from code that is not holding
780 * cgroup_mutex (not nested in a cgroup_lock() call.)
782 void rebuild_sched_domains(void)
784 do_rebuild_sched_domains(NULL);
788 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
790 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
792 * Call with cgroup_mutex held. May take callback_mutex during call.
793 * Called for each task in a cgroup by cgroup_scan_tasks().
794 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
795 * words, if its mask is not equal to its cpuset's mask).
797 static int cpuset_test_cpumask(struct task_struct *tsk,
798 struct cgroup_scanner *scan)
800 return !cpumask_equal(&tsk->cpus_allowed,
801 (cgroup_cs(scan->cg))->cpus_allowed);
805 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
807 * @scan: struct cgroup_scanner containing the cgroup of the task
809 * Called by cgroup_scan_tasks() for each task in a cgroup whose
810 * cpus_allowed mask needs to be changed.
812 * We don't need to re-check for the cgroup/cpuset membership, since we're
813 * holding cgroup_lock() at this point.
815 static void cpuset_change_cpumask(struct task_struct *tsk,
816 struct cgroup_scanner *scan)
818 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
822 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
823 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
824 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
826 * Called with cgroup_mutex held
828 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
829 * calling callback functions for each.
831 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
834 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
836 struct cgroup_scanner scan;
838 scan.cg = cs->css.cgroup;
839 scan.test_task = cpuset_test_cpumask;
840 scan.process_task = cpuset_change_cpumask;
842 cgroup_scan_tasks(&scan);
846 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
847 * @cs: the cpuset to consider
848 * @buf: buffer of cpu numbers written to this cpuset
850 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
853 struct ptr_heap heap;
855 int is_load_balanced;
857 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
858 if (cs == &top_cpuset)
862 * An empty cpus_allowed is ok only if the cpuset has no tasks.
863 * Since cpulist_parse() fails on an empty mask, we special case
864 * that parsing. The validate_change() call ensures that cpusets
865 * with tasks have cpus.
868 cpumask_clear(trialcs->cpus_allowed);
870 retval = cpulist_parse(buf, trialcs->cpus_allowed);
874 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
877 retval = validate_change(cs, trialcs);
881 /* Nothing to do if the cpus didn't change */
882 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
885 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
889 is_load_balanced = is_sched_load_balance(trialcs);
891 mutex_lock(&callback_mutex);
892 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
893 mutex_unlock(&callback_mutex);
896 * Scan tasks in the cpuset, and update the cpumasks of any
897 * that need an update.
899 update_tasks_cpumask(cs, &heap);
903 if (is_load_balanced)
904 async_rebuild_sched_domains();
911 * Migrate memory region from one set of nodes to another.
913 * Temporarilly set tasks mems_allowed to target nodes of migration,
914 * so that the migration code can allocate pages on these nodes.
916 * Call holding cgroup_mutex, so current's cpuset won't change
917 * during this call, as manage_mutex holds off any cpuset_attach()
918 * calls. Therefore we don't need to take task_lock around the
919 * call to guarantee_online_mems(), as we know no one is changing
922 * While the mm_struct we are migrating is typically from some
923 * other task, the task_struct mems_allowed that we are hacking
924 * is for our current task, which must allocate new pages for that
925 * migrating memory region.
928 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
929 const nodemask_t *to)
931 struct task_struct *tsk = current;
933 tsk->mems_allowed = *to;
935 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
937 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
941 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
942 * @tsk: the task to change
943 * @newmems: new nodes that the task will be set
945 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
946 * we structure updates as setting all new allowed nodes, then clearing newly
949 static void cpuset_change_task_nodemask(struct task_struct *tsk,
954 * Allow tasks that have access to memory reserves because they have
955 * been OOM killed to get memory anywhere.
957 if (unlikely(test_thread_flag(TIF_MEMDIE)))
959 if (current->flags & PF_EXITING) /* Let dying task have memory */
963 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
964 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
968 * ensure checking ->mems_allowed_change_disable after setting all new
971 * the read-side task can see an nodemask with new allowed nodes and
972 * old allowed nodes. and if it allocates page when cpuset clears newly
973 * disallowed ones continuous, it can see the new allowed bits.
975 * And if setting all new allowed nodes is after the checking, setting
976 * all new allowed nodes and clearing newly disallowed ones will be done
977 * continuous, and the read-side task may find no node to alloc page.
982 * Allocation of memory is very fast, we needn't sleep when waiting
985 while (ACCESS_ONCE(tsk->mems_allowed_change_disable)) {
993 * ensure checking ->mems_allowed_change_disable before clearing all new
996 * if clearing newly disallowed bits before the checking, the read-side
997 * task may find no node to alloc page.
1001 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1002 tsk->mems_allowed = *newmems;
1007 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1008 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1009 * memory_migrate flag is set. Called with cgroup_mutex held.
1011 static void cpuset_change_nodemask(struct task_struct *p,
1012 struct cgroup_scanner *scan)
1014 struct mm_struct *mm;
1017 const nodemask_t *oldmem = scan->data;
1018 NODEMASK_ALLOC(nodemask_t, newmems, GFP_KERNEL);
1023 cs = cgroup_cs(scan->cg);
1024 guarantee_online_mems(cs, newmems);
1026 cpuset_change_task_nodemask(p, newmems);
1028 NODEMASK_FREE(newmems);
1030 mm = get_task_mm(p);
1034 migrate = is_memory_migrate(cs);
1036 mpol_rebind_mm(mm, &cs->mems_allowed);
1038 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1042 static void *cpuset_being_rebound;
1045 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1046 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1047 * @oldmem: old mems_allowed of cpuset cs
1048 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1050 * Called with cgroup_mutex held
1051 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1054 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1055 struct ptr_heap *heap)
1057 struct cgroup_scanner scan;
1059 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1061 scan.cg = cs->css.cgroup;
1062 scan.test_task = NULL;
1063 scan.process_task = cpuset_change_nodemask;
1065 scan.data = (nodemask_t *)oldmem;
1068 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1069 * take while holding tasklist_lock. Forks can happen - the
1070 * mpol_dup() cpuset_being_rebound check will catch such forks,
1071 * and rebind their vma mempolicies too. Because we still hold
1072 * the global cgroup_mutex, we know that no other rebind effort
1073 * will be contending for the global variable cpuset_being_rebound.
1074 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1075 * is idempotent. Also migrate pages in each mm to new nodes.
1077 cgroup_scan_tasks(&scan);
1079 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1080 cpuset_being_rebound = NULL;
1084 * Handle user request to change the 'mems' memory placement
1085 * of a cpuset. Needs to validate the request, update the
1086 * cpusets mems_allowed, and for each task in the cpuset,
1087 * update mems_allowed and rebind task's mempolicy and any vma
1088 * mempolicies and if the cpuset is marked 'memory_migrate',
1089 * migrate the tasks pages to the new memory.
1091 * Call with cgroup_mutex held. May take callback_mutex during call.
1092 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1093 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1094 * their mempolicies to the cpusets new mems_allowed.
1096 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1099 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1101 struct ptr_heap heap;
1107 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1110 if (cs == &top_cpuset) {
1116 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1117 * Since nodelist_parse() fails on an empty mask, we special case
1118 * that parsing. The validate_change() call ensures that cpusets
1119 * with tasks have memory.
1122 nodes_clear(trialcs->mems_allowed);
1124 retval = nodelist_parse(buf, trialcs->mems_allowed);
1128 if (!nodes_subset(trialcs->mems_allowed,
1129 node_states[N_HIGH_MEMORY])) {
1134 *oldmem = cs->mems_allowed;
1135 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1136 retval = 0; /* Too easy - nothing to do */
1139 retval = validate_change(cs, trialcs);
1143 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1147 mutex_lock(&callback_mutex);
1148 cs->mems_allowed = trialcs->mems_allowed;
1149 mutex_unlock(&callback_mutex);
1151 update_tasks_nodemask(cs, oldmem, &heap);
1155 NODEMASK_FREE(oldmem);
1159 int current_cpuset_is_being_rebound(void)
1161 return task_cs(current) == cpuset_being_rebound;
1164 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1167 if (val < -1 || val >= SD_LV_MAX)
1171 if (val != cs->relax_domain_level) {
1172 cs->relax_domain_level = val;
1173 if (!cpumask_empty(cs->cpus_allowed) &&
1174 is_sched_load_balance(cs))
1175 async_rebuild_sched_domains();
1182 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1183 * @tsk: task to be updated
1184 * @scan: struct cgroup_scanner containing the cgroup of the task
1186 * Called by cgroup_scan_tasks() for each task in a cgroup.
1188 * We don't need to re-check for the cgroup/cpuset membership, since we're
1189 * holding cgroup_lock() at this point.
1191 static void cpuset_change_flag(struct task_struct *tsk,
1192 struct cgroup_scanner *scan)
1194 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1198 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1199 * @cs: the cpuset in which each task's spread flags needs to be changed
1200 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1202 * Called with cgroup_mutex held
1204 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1205 * calling callback functions for each.
1207 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1210 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1212 struct cgroup_scanner scan;
1214 scan.cg = cs->css.cgroup;
1215 scan.test_task = NULL;
1216 scan.process_task = cpuset_change_flag;
1218 cgroup_scan_tasks(&scan);
1222 * update_flag - read a 0 or a 1 in a file and update associated flag
1223 * bit: the bit to update (see cpuset_flagbits_t)
1224 * cs: the cpuset to update
1225 * turning_on: whether the flag is being set or cleared
1227 * Call with cgroup_mutex held.
1230 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1233 struct cpuset *trialcs;
1234 int balance_flag_changed;
1235 int spread_flag_changed;
1236 struct ptr_heap heap;
1239 trialcs = alloc_trial_cpuset(cs);
1244 set_bit(bit, &trialcs->flags);
1246 clear_bit(bit, &trialcs->flags);
1248 err = validate_change(cs, trialcs);
1252 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1256 balance_flag_changed = (is_sched_load_balance(cs) !=
1257 is_sched_load_balance(trialcs));
1259 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1260 || (is_spread_page(cs) != is_spread_page(trialcs)));
1262 mutex_lock(&callback_mutex);
1263 cs->flags = trialcs->flags;
1264 mutex_unlock(&callback_mutex);
1266 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1267 async_rebuild_sched_domains();
1269 if (spread_flag_changed)
1270 update_tasks_flags(cs, &heap);
1273 free_trial_cpuset(trialcs);
1278 * Frequency meter - How fast is some event occurring?
1280 * These routines manage a digitally filtered, constant time based,
1281 * event frequency meter. There are four routines:
1282 * fmeter_init() - initialize a frequency meter.
1283 * fmeter_markevent() - called each time the event happens.
1284 * fmeter_getrate() - returns the recent rate of such events.
1285 * fmeter_update() - internal routine used to update fmeter.
1287 * A common data structure is passed to each of these routines,
1288 * which is used to keep track of the state required to manage the
1289 * frequency meter and its digital filter.
1291 * The filter works on the number of events marked per unit time.
1292 * The filter is single-pole low-pass recursive (IIR). The time unit
1293 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1294 * simulate 3 decimal digits of precision (multiplied by 1000).
1296 * With an FM_COEF of 933, and a time base of 1 second, the filter
1297 * has a half-life of 10 seconds, meaning that if the events quit
1298 * happening, then the rate returned from the fmeter_getrate()
1299 * will be cut in half each 10 seconds, until it converges to zero.
1301 * It is not worth doing a real infinitely recursive filter. If more
1302 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1303 * just compute FM_MAXTICKS ticks worth, by which point the level
1306 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1307 * arithmetic overflow in the fmeter_update() routine.
1309 * Given the simple 32 bit integer arithmetic used, this meter works
1310 * best for reporting rates between one per millisecond (msec) and
1311 * one per 32 (approx) seconds. At constant rates faster than one
1312 * per msec it maxes out at values just under 1,000,000. At constant
1313 * rates between one per msec, and one per second it will stabilize
1314 * to a value N*1000, where N is the rate of events per second.
1315 * At constant rates between one per second and one per 32 seconds,
1316 * it will be choppy, moving up on the seconds that have an event,
1317 * and then decaying until the next event. At rates slower than
1318 * about one in 32 seconds, it decays all the way back to zero between
1322 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1323 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1324 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1325 #define FM_SCALE 1000 /* faux fixed point scale */
1327 /* Initialize a frequency meter */
1328 static void fmeter_init(struct fmeter *fmp)
1333 spin_lock_init(&fmp->lock);
1336 /* Internal meter update - process cnt events and update value */
1337 static void fmeter_update(struct fmeter *fmp)
1339 time_t now = get_seconds();
1340 time_t ticks = now - fmp->time;
1345 ticks = min(FM_MAXTICKS, ticks);
1347 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1350 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1354 /* Process any previous ticks, then bump cnt by one (times scale). */
1355 static void fmeter_markevent(struct fmeter *fmp)
1357 spin_lock(&fmp->lock);
1359 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1360 spin_unlock(&fmp->lock);
1363 /* Process any previous ticks, then return current value. */
1364 static int fmeter_getrate(struct fmeter *fmp)
1368 spin_lock(&fmp->lock);
1371 spin_unlock(&fmp->lock);
1375 /* Protected by cgroup_lock */
1376 static cpumask_var_t cpus_attach;
1378 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1379 static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1380 struct task_struct *tsk, bool threadgroup)
1383 struct cpuset *cs = cgroup_cs(cont);
1385 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1389 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1390 * cannot change their cpu affinity and isolating such threads by their
1391 * set of allowed nodes is unnecessary. Thus, cpusets are not
1392 * applicable for such threads. This prevents checking for success of
1393 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1396 if (tsk->flags & PF_THREAD_BOUND)
1399 ret = security_task_setscheduler(tsk);
1403 struct task_struct *c;
1406 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
1407 ret = security_task_setscheduler(c);
1418 static void cpuset_attach_task(struct task_struct *tsk, nodemask_t *to,
1423 * can_attach beforehand should guarantee that this doesn't fail.
1424 * TODO: have a better way to handle failure here
1426 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1429 cpuset_change_task_nodemask(tsk, to);
1430 cpuset_update_task_spread_flag(cs, tsk);
1434 static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1435 struct cgroup *oldcont, struct task_struct *tsk,
1438 struct mm_struct *mm;
1439 struct cpuset *cs = cgroup_cs(cont);
1440 struct cpuset *oldcs = cgroup_cs(oldcont);
1441 NODEMASK_ALLOC(nodemask_t, to, GFP_KERNEL);
1446 if (cs == &top_cpuset) {
1447 cpumask_copy(cpus_attach, cpu_possible_mask);
1449 guarantee_online_cpus(cs, cpus_attach);
1451 guarantee_online_mems(cs, to);
1453 /* do per-task migration stuff possibly for each in the threadgroup */
1454 cpuset_attach_task(tsk, to, cs);
1456 struct task_struct *c;
1458 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
1459 cpuset_attach_task(c, to, cs);
1464 /* change mm; only needs to be done once even if threadgroup */
1465 *to = cs->mems_allowed;
1466 mm = get_task_mm(tsk);
1468 mpol_rebind_mm(mm, to);
1469 if (is_memory_migrate(cs))
1470 cpuset_migrate_mm(mm, &oldcs->mems_allowed, to);
1478 /* The various types of files and directories in a cpuset file system */
1481 FILE_MEMORY_MIGRATE,
1487 FILE_SCHED_LOAD_BALANCE,
1488 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1489 FILE_MEMORY_PRESSURE_ENABLED,
1490 FILE_MEMORY_PRESSURE,
1493 } cpuset_filetype_t;
1495 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1498 struct cpuset *cs = cgroup_cs(cgrp);
1499 cpuset_filetype_t type = cft->private;
1501 if (!cgroup_lock_live_group(cgrp))
1505 case FILE_CPU_EXCLUSIVE:
1506 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1508 case FILE_MEM_EXCLUSIVE:
1509 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1511 case FILE_MEM_HARDWALL:
1512 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1514 case FILE_SCHED_LOAD_BALANCE:
1515 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1517 case FILE_MEMORY_MIGRATE:
1518 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1520 case FILE_MEMORY_PRESSURE_ENABLED:
1521 cpuset_memory_pressure_enabled = !!val;
1523 case FILE_MEMORY_PRESSURE:
1526 case FILE_SPREAD_PAGE:
1527 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1529 case FILE_SPREAD_SLAB:
1530 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1540 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1543 struct cpuset *cs = cgroup_cs(cgrp);
1544 cpuset_filetype_t type = cft->private;
1546 if (!cgroup_lock_live_group(cgrp))
1550 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1551 retval = update_relax_domain_level(cs, val);
1562 * Common handling for a write to a "cpus" or "mems" file.
1564 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1568 struct cpuset *cs = cgroup_cs(cgrp);
1569 struct cpuset *trialcs;
1571 if (!cgroup_lock_live_group(cgrp))
1574 trialcs = alloc_trial_cpuset(cs);
1580 switch (cft->private) {
1582 retval = update_cpumask(cs, trialcs, buf);
1585 retval = update_nodemask(cs, trialcs, buf);
1592 free_trial_cpuset(trialcs);
1599 * These ascii lists should be read in a single call, by using a user
1600 * buffer large enough to hold the entire map. If read in smaller
1601 * chunks, there is no guarantee of atomicity. Since the display format
1602 * used, list of ranges of sequential numbers, is variable length,
1603 * and since these maps can change value dynamically, one could read
1604 * gibberish by doing partial reads while a list was changing.
1605 * A single large read to a buffer that crosses a page boundary is
1606 * ok, because the result being copied to user land is not recomputed
1607 * across a page fault.
1610 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1614 mutex_lock(&callback_mutex);
1615 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1616 mutex_unlock(&callback_mutex);
1621 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1625 mutex_lock(&callback_mutex);
1626 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1627 mutex_unlock(&callback_mutex);
1632 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1636 size_t nbytes, loff_t *ppos)
1638 struct cpuset *cs = cgroup_cs(cont);
1639 cpuset_filetype_t type = cft->private;
1644 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1651 s += cpuset_sprintf_cpulist(s, cs);
1654 s += cpuset_sprintf_memlist(s, cs);
1662 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1664 free_page((unsigned long)page);
1668 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1670 struct cpuset *cs = cgroup_cs(cont);
1671 cpuset_filetype_t type = cft->private;
1673 case FILE_CPU_EXCLUSIVE:
1674 return is_cpu_exclusive(cs);
1675 case FILE_MEM_EXCLUSIVE:
1676 return is_mem_exclusive(cs);
1677 case FILE_MEM_HARDWALL:
1678 return is_mem_hardwall(cs);
1679 case FILE_SCHED_LOAD_BALANCE:
1680 return is_sched_load_balance(cs);
1681 case FILE_MEMORY_MIGRATE:
1682 return is_memory_migrate(cs);
1683 case FILE_MEMORY_PRESSURE_ENABLED:
1684 return cpuset_memory_pressure_enabled;
1685 case FILE_MEMORY_PRESSURE:
1686 return fmeter_getrate(&cs->fmeter);
1687 case FILE_SPREAD_PAGE:
1688 return is_spread_page(cs);
1689 case FILE_SPREAD_SLAB:
1690 return is_spread_slab(cs);
1695 /* Unreachable but makes gcc happy */
1699 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1701 struct cpuset *cs = cgroup_cs(cont);
1702 cpuset_filetype_t type = cft->private;
1704 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1705 return cs->relax_domain_level;
1710 /* Unrechable but makes gcc happy */
1716 * for the common functions, 'private' gives the type of file
1719 static struct cftype files[] = {
1722 .read = cpuset_common_file_read,
1723 .write_string = cpuset_write_resmask,
1724 .max_write_len = (100U + 6 * NR_CPUS),
1725 .private = FILE_CPULIST,
1730 .read = cpuset_common_file_read,
1731 .write_string = cpuset_write_resmask,
1732 .max_write_len = (100U + 6 * MAX_NUMNODES),
1733 .private = FILE_MEMLIST,
1737 .name = "cpu_exclusive",
1738 .read_u64 = cpuset_read_u64,
1739 .write_u64 = cpuset_write_u64,
1740 .private = FILE_CPU_EXCLUSIVE,
1744 .name = "mem_exclusive",
1745 .read_u64 = cpuset_read_u64,
1746 .write_u64 = cpuset_write_u64,
1747 .private = FILE_MEM_EXCLUSIVE,
1751 .name = "mem_hardwall",
1752 .read_u64 = cpuset_read_u64,
1753 .write_u64 = cpuset_write_u64,
1754 .private = FILE_MEM_HARDWALL,
1758 .name = "sched_load_balance",
1759 .read_u64 = cpuset_read_u64,
1760 .write_u64 = cpuset_write_u64,
1761 .private = FILE_SCHED_LOAD_BALANCE,
1765 .name = "sched_relax_domain_level",
1766 .read_s64 = cpuset_read_s64,
1767 .write_s64 = cpuset_write_s64,
1768 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1772 .name = "memory_migrate",
1773 .read_u64 = cpuset_read_u64,
1774 .write_u64 = cpuset_write_u64,
1775 .private = FILE_MEMORY_MIGRATE,
1779 .name = "memory_pressure",
1780 .read_u64 = cpuset_read_u64,
1781 .write_u64 = cpuset_write_u64,
1782 .private = FILE_MEMORY_PRESSURE,
1787 .name = "memory_spread_page",
1788 .read_u64 = cpuset_read_u64,
1789 .write_u64 = cpuset_write_u64,
1790 .private = FILE_SPREAD_PAGE,
1794 .name = "memory_spread_slab",
1795 .read_u64 = cpuset_read_u64,
1796 .write_u64 = cpuset_write_u64,
1797 .private = FILE_SPREAD_SLAB,
1801 static struct cftype cft_memory_pressure_enabled = {
1802 .name = "memory_pressure_enabled",
1803 .read_u64 = cpuset_read_u64,
1804 .write_u64 = cpuset_write_u64,
1805 .private = FILE_MEMORY_PRESSURE_ENABLED,
1808 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1812 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1815 /* memory_pressure_enabled is in root cpuset only */
1817 err = cgroup_add_file(cont, ss,
1818 &cft_memory_pressure_enabled);
1823 * post_clone() is called at the end of cgroup_clone().
1824 * 'cgroup' was just created automatically as a result of
1825 * a cgroup_clone(), and the current task is about to
1826 * be moved into 'cgroup'.
1828 * Currently we refuse to set up the cgroup - thereby
1829 * refusing the task to be entered, and as a result refusing
1830 * the sys_unshare() or clone() which initiated it - if any
1831 * sibling cpusets have exclusive cpus or mem.
1833 * If this becomes a problem for some users who wish to
1834 * allow that scenario, then cpuset_post_clone() could be
1835 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1836 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1839 static void cpuset_post_clone(struct cgroup_subsys *ss,
1840 struct cgroup *cgroup)
1842 struct cgroup *parent, *child;
1843 struct cpuset *cs, *parent_cs;
1845 parent = cgroup->parent;
1846 list_for_each_entry(child, &parent->children, sibling) {
1847 cs = cgroup_cs(child);
1848 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1851 cs = cgroup_cs(cgroup);
1852 parent_cs = cgroup_cs(parent);
1854 cs->mems_allowed = parent_cs->mems_allowed;
1855 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1860 * cpuset_create - create a cpuset
1861 * ss: cpuset cgroup subsystem
1862 * cont: control group that the new cpuset will be part of
1865 static struct cgroup_subsys_state *cpuset_create(
1866 struct cgroup_subsys *ss,
1867 struct cgroup *cont)
1870 struct cpuset *parent;
1872 if (!cont->parent) {
1873 return &top_cpuset.css;
1875 parent = cgroup_cs(cont->parent);
1876 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1878 return ERR_PTR(-ENOMEM);
1879 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1881 return ERR_PTR(-ENOMEM);
1885 if (is_spread_page(parent))
1886 set_bit(CS_SPREAD_PAGE, &cs->flags);
1887 if (is_spread_slab(parent))
1888 set_bit(CS_SPREAD_SLAB, &cs->flags);
1889 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1890 cpumask_clear(cs->cpus_allowed);
1891 nodes_clear(cs->mems_allowed);
1892 fmeter_init(&cs->fmeter);
1893 cs->relax_domain_level = -1;
1895 cs->parent = parent;
1896 number_of_cpusets++;
1901 * If the cpuset being removed has its flag 'sched_load_balance'
1902 * enabled, then simulate turning sched_load_balance off, which
1903 * will call async_rebuild_sched_domains().
1906 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1908 struct cpuset *cs = cgroup_cs(cont);
1910 if (is_sched_load_balance(cs))
1911 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1913 number_of_cpusets--;
1914 free_cpumask_var(cs->cpus_allowed);
1918 struct cgroup_subsys cpuset_subsys = {
1920 .create = cpuset_create,
1921 .destroy = cpuset_destroy,
1922 .can_attach = cpuset_can_attach,
1923 .attach = cpuset_attach,
1924 .populate = cpuset_populate,
1925 .post_clone = cpuset_post_clone,
1926 .subsys_id = cpuset_subsys_id,
1931 * cpuset_init - initialize cpusets at system boot
1933 * Description: Initialize top_cpuset and the cpuset internal file system,
1936 int __init cpuset_init(void)
1940 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1943 cpumask_setall(top_cpuset.cpus_allowed);
1944 nodes_setall(top_cpuset.mems_allowed);
1946 fmeter_init(&top_cpuset.fmeter);
1947 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1948 top_cpuset.relax_domain_level = -1;
1950 err = register_filesystem(&cpuset_fs_type);
1954 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1957 number_of_cpusets = 1;
1962 * cpuset_do_move_task - move a given task to another cpuset
1963 * @tsk: pointer to task_struct the task to move
1964 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1966 * Called by cgroup_scan_tasks() for each task in a cgroup.
1967 * Return nonzero to stop the walk through the tasks.
1969 static void cpuset_do_move_task(struct task_struct *tsk,
1970 struct cgroup_scanner *scan)
1972 struct cgroup *new_cgroup = scan->data;
1974 cgroup_attach_task(new_cgroup, tsk);
1978 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1979 * @from: cpuset in which the tasks currently reside
1980 * @to: cpuset to which the tasks will be moved
1982 * Called with cgroup_mutex held
1983 * callback_mutex must not be held, as cpuset_attach() will take it.
1985 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1986 * calling callback functions for each.
1988 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1990 struct cgroup_scanner scan;
1992 scan.cg = from->css.cgroup;
1993 scan.test_task = NULL; /* select all tasks in cgroup */
1994 scan.process_task = cpuset_do_move_task;
1996 scan.data = to->css.cgroup;
1998 if (cgroup_scan_tasks(&scan))
1999 printk(KERN_ERR "move_member_tasks_to_cpuset: "
2000 "cgroup_scan_tasks failed\n");
2004 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2005 * or memory nodes, we need to walk over the cpuset hierarchy,
2006 * removing that CPU or node from all cpusets. If this removes the
2007 * last CPU or node from a cpuset, then move the tasks in the empty
2008 * cpuset to its next-highest non-empty parent.
2010 * Called with cgroup_mutex held
2011 * callback_mutex must not be held, as cpuset_attach() will take it.
2013 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2015 struct cpuset *parent;
2018 * The cgroup's css_sets list is in use if there are tasks
2019 * in the cpuset; the list is empty if there are none;
2020 * the cs->css.refcnt seems always 0.
2022 if (list_empty(&cs->css.cgroup->css_sets))
2026 * Find its next-highest non-empty parent, (top cpuset
2027 * has online cpus, so can't be empty).
2029 parent = cs->parent;
2030 while (cpumask_empty(parent->cpus_allowed) ||
2031 nodes_empty(parent->mems_allowed))
2032 parent = parent->parent;
2034 move_member_tasks_to_cpuset(cs, parent);
2038 * Walk the specified cpuset subtree and look for empty cpusets.
2039 * The tasks of such cpuset must be moved to a parent cpuset.
2041 * Called with cgroup_mutex held. We take callback_mutex to modify
2042 * cpus_allowed and mems_allowed.
2044 * This walk processes the tree from top to bottom, completing one layer
2045 * before dropping down to the next. It always processes a node before
2046 * any of its children.
2048 * For now, since we lack memory hot unplug, we'll never see a cpuset
2049 * that has tasks along with an empty 'mems'. But if we did see such
2050 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2052 static void scan_for_empty_cpusets(struct cpuset *root)
2055 struct cpuset *cp; /* scans cpusets being updated */
2056 struct cpuset *child; /* scans child cpusets of cp */
2057 struct cgroup *cont;
2058 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2060 if (oldmems == NULL)
2063 list_add_tail((struct list_head *)&root->stack_list, &queue);
2065 while (!list_empty(&queue)) {
2066 cp = list_first_entry(&queue, struct cpuset, stack_list);
2067 list_del(queue.next);
2068 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2069 child = cgroup_cs(cont);
2070 list_add_tail(&child->stack_list, &queue);
2073 /* Continue past cpusets with all cpus, mems online */
2074 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
2075 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2078 *oldmems = cp->mems_allowed;
2080 /* Remove offline cpus and mems from this cpuset. */
2081 mutex_lock(&callback_mutex);
2082 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2084 nodes_and(cp->mems_allowed, cp->mems_allowed,
2085 node_states[N_HIGH_MEMORY]);
2086 mutex_unlock(&callback_mutex);
2088 /* Move tasks from the empty cpuset to a parent */
2089 if (cpumask_empty(cp->cpus_allowed) ||
2090 nodes_empty(cp->mems_allowed))
2091 remove_tasks_in_empty_cpuset(cp);
2093 update_tasks_cpumask(cp, NULL);
2094 update_tasks_nodemask(cp, oldmems, NULL);
2097 NODEMASK_FREE(oldmems);
2101 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2102 * period. This is necessary in order to make cpusets transparent
2103 * (of no affect) on systems that are actively using CPU hotplug
2104 * but making no active use of cpusets.
2106 * This routine ensures that top_cpuset.cpus_allowed tracks
2107 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2109 * Called within get_online_cpus(). Needs to call cgroup_lock()
2110 * before calling generate_sched_domains().
2112 void cpuset_update_active_cpus(void)
2114 struct sched_domain_attr *attr;
2115 cpumask_var_t *doms;
2119 mutex_lock(&callback_mutex);
2120 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2121 mutex_unlock(&callback_mutex);
2122 scan_for_empty_cpusets(&top_cpuset);
2123 ndoms = generate_sched_domains(&doms, &attr);
2126 /* Have scheduler rebuild the domains */
2127 partition_sched_domains(ndoms, doms, attr);
2130 #ifdef CONFIG_MEMORY_HOTPLUG
2132 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2133 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2134 * See also the previous routine cpuset_track_online_cpus().
2136 static int cpuset_track_online_nodes(struct notifier_block *self,
2137 unsigned long action, void *arg)
2139 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2141 if (oldmems == NULL)
2147 *oldmems = top_cpuset.mems_allowed;
2148 mutex_lock(&callback_mutex);
2149 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2150 mutex_unlock(&callback_mutex);
2151 update_tasks_nodemask(&top_cpuset, oldmems, NULL);
2155 * needn't update top_cpuset.mems_allowed explicitly because
2156 * scan_for_empty_cpusets() will update it.
2158 scan_for_empty_cpusets(&top_cpuset);
2165 NODEMASK_FREE(oldmems);
2171 * cpuset_init_smp - initialize cpus_allowed
2173 * Description: Finish top cpuset after cpu, node maps are initialized
2176 void __init cpuset_init_smp(void)
2178 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2179 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2181 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2183 cpuset_wq = create_singlethread_workqueue("cpuset");
2188 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2189 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2190 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2192 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2193 * attached to the specified @tsk. Guaranteed to return some non-empty
2194 * subset of cpu_online_map, even if this means going outside the
2198 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2200 mutex_lock(&callback_mutex);
2202 guarantee_online_cpus(task_cs(tsk), pmask);
2204 mutex_unlock(&callback_mutex);
2207 int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2209 const struct cpuset *cs;
2215 cpumask_copy(&tsk->cpus_allowed, cs->cpus_allowed);
2219 * We own tsk->cpus_allowed, nobody can change it under us.
2221 * But we used cs && cs->cpus_allowed lockless and thus can
2222 * race with cgroup_attach_task() or update_cpumask() and get
2223 * the wrong tsk->cpus_allowed. However, both cases imply the
2224 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2225 * which takes task_rq_lock().
2227 * If we are called after it dropped the lock we must see all
2228 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2229 * set any mask even if it is not right from task_cs() pov,
2230 * the pending set_cpus_allowed_ptr() will fix things.
2233 cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2234 if (cpu >= nr_cpu_ids) {
2236 * Either tsk->cpus_allowed is wrong (see above) or it
2237 * is actually empty. The latter case is only possible
2238 * if we are racing with remove_tasks_in_empty_cpuset().
2239 * Like above we can temporary set any mask and rely on
2240 * set_cpus_allowed_ptr() as synchronization point.
2242 cpumask_copy(&tsk->cpus_allowed, cpu_possible_mask);
2243 cpu = cpumask_any(cpu_active_mask);
2249 void cpuset_init_current_mems_allowed(void)
2251 nodes_setall(current->mems_allowed);
2255 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2256 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2258 * Description: Returns the nodemask_t mems_allowed of the cpuset
2259 * attached to the specified @tsk. Guaranteed to return some non-empty
2260 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2264 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2268 mutex_lock(&callback_mutex);
2270 guarantee_online_mems(task_cs(tsk), &mask);
2272 mutex_unlock(&callback_mutex);
2278 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2279 * @nodemask: the nodemask to be checked
2281 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2283 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2285 return nodes_intersects(*nodemask, current->mems_allowed);
2289 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2290 * mem_hardwall ancestor to the specified cpuset. Call holding
2291 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2292 * (an unusual configuration), then returns the root cpuset.
2294 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2296 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2302 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2303 * @node: is this an allowed node?
2304 * @gfp_mask: memory allocation flags
2306 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2307 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2308 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2309 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2310 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2314 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2315 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2316 * might sleep, and might allow a node from an enclosing cpuset.
2318 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2319 * cpusets, and never sleeps.
2321 * The __GFP_THISNODE placement logic is really handled elsewhere,
2322 * by forcibly using a zonelist starting at a specified node, and by
2323 * (in get_page_from_freelist()) refusing to consider the zones for
2324 * any node on the zonelist except the first. By the time any such
2325 * calls get to this routine, we should just shut up and say 'yes'.
2327 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2328 * and do not allow allocations outside the current tasks cpuset
2329 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2330 * GFP_KERNEL allocations are not so marked, so can escape to the
2331 * nearest enclosing hardwalled ancestor cpuset.
2333 * Scanning up parent cpusets requires callback_mutex. The
2334 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2335 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2336 * current tasks mems_allowed came up empty on the first pass over
2337 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2338 * cpuset are short of memory, might require taking the callback_mutex
2341 * The first call here from mm/page_alloc:get_page_from_freelist()
2342 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2343 * so no allocation on a node outside the cpuset is allowed (unless
2344 * in interrupt, of course).
2346 * The second pass through get_page_from_freelist() doesn't even call
2347 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2348 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2349 * in alloc_flags. That logic and the checks below have the combined
2351 * in_interrupt - any node ok (current task context irrelevant)
2352 * GFP_ATOMIC - any node ok
2353 * TIF_MEMDIE - any node ok
2354 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2355 * GFP_USER - only nodes in current tasks mems allowed ok.
2358 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2359 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2360 * the code that might scan up ancestor cpusets and sleep.
2362 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2364 const struct cpuset *cs; /* current cpuset ancestors */
2365 int allowed; /* is allocation in zone z allowed? */
2367 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2369 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2370 if (node_isset(node, current->mems_allowed))
2373 * Allow tasks that have access to memory reserves because they have
2374 * been OOM killed to get memory anywhere.
2376 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2378 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2381 if (current->flags & PF_EXITING) /* Let dying task have memory */
2384 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2385 mutex_lock(&callback_mutex);
2388 cs = nearest_hardwall_ancestor(task_cs(current));
2389 task_unlock(current);
2391 allowed = node_isset(node, cs->mems_allowed);
2392 mutex_unlock(&callback_mutex);
2397 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2398 * @node: is this an allowed node?
2399 * @gfp_mask: memory allocation flags
2401 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2402 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2403 * yes. If the task has been OOM killed and has access to memory reserves as
2404 * specified by the TIF_MEMDIE flag, yes.
2407 * The __GFP_THISNODE placement logic is really handled elsewhere,
2408 * by forcibly using a zonelist starting at a specified node, and by
2409 * (in get_page_from_freelist()) refusing to consider the zones for
2410 * any node on the zonelist except the first. By the time any such
2411 * calls get to this routine, we should just shut up and say 'yes'.
2413 * Unlike the cpuset_node_allowed_softwall() variant, above,
2414 * this variant requires that the node be in the current task's
2415 * mems_allowed or that we're in interrupt. It does not scan up the
2416 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2419 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2421 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2423 if (node_isset(node, current->mems_allowed))
2426 * Allow tasks that have access to memory reserves because they have
2427 * been OOM killed to get memory anywhere.
2429 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2435 * cpuset_unlock - release lock on cpuset changes
2437 * Undo the lock taken in a previous cpuset_lock() call.
2440 void cpuset_unlock(void)
2442 mutex_unlock(&callback_mutex);
2446 * cpuset_mem_spread_node() - On which node to begin search for a file page
2447 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2449 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2450 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2451 * and if the memory allocation used cpuset_mem_spread_node()
2452 * to determine on which node to start looking, as it will for
2453 * certain page cache or slab cache pages such as used for file
2454 * system buffers and inode caches, then instead of starting on the
2455 * local node to look for a free page, rather spread the starting
2456 * node around the tasks mems_allowed nodes.
2458 * We don't have to worry about the returned node being offline
2459 * because "it can't happen", and even if it did, it would be ok.
2461 * The routines calling guarantee_online_mems() are careful to
2462 * only set nodes in task->mems_allowed that are online. So it
2463 * should not be possible for the following code to return an
2464 * offline node. But if it did, that would be ok, as this routine
2465 * is not returning the node where the allocation must be, only
2466 * the node where the search should start. The zonelist passed to
2467 * __alloc_pages() will include all nodes. If the slab allocator
2468 * is passed an offline node, it will fall back to the local node.
2469 * See kmem_cache_alloc_node().
2472 static int cpuset_spread_node(int *rotor)
2476 node = next_node(*rotor, current->mems_allowed);
2477 if (node == MAX_NUMNODES)
2478 node = first_node(current->mems_allowed);
2483 int cpuset_mem_spread_node(void)
2485 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2488 int cpuset_slab_spread_node(void)
2490 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2493 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2496 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2497 * @tsk1: pointer to task_struct of some task.
2498 * @tsk2: pointer to task_struct of some other task.
2500 * Description: Return true if @tsk1's mems_allowed intersects the
2501 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2502 * one of the task's memory usage might impact the memory available
2506 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2507 const struct task_struct *tsk2)
2509 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2513 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2514 * @task: pointer to task_struct of some task.
2516 * Description: Prints @task's name, cpuset name, and cached copy of its
2517 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2518 * dereferencing task_cs(task).
2520 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2522 struct dentry *dentry;
2524 dentry = task_cs(tsk)->css.cgroup->dentry;
2525 spin_lock(&cpuset_buffer_lock);
2526 snprintf(cpuset_name, CPUSET_NAME_LEN,
2527 dentry ? (const char *)dentry->d_name.name : "/");
2528 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2530 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2531 tsk->comm, cpuset_name, cpuset_nodelist);
2532 spin_unlock(&cpuset_buffer_lock);
2536 * Collection of memory_pressure is suppressed unless
2537 * this flag is enabled by writing "1" to the special
2538 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2541 int cpuset_memory_pressure_enabled __read_mostly;
2544 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2546 * Keep a running average of the rate of synchronous (direct)
2547 * page reclaim efforts initiated by tasks in each cpuset.
2549 * This represents the rate at which some task in the cpuset
2550 * ran low on memory on all nodes it was allowed to use, and
2551 * had to enter the kernels page reclaim code in an effort to
2552 * create more free memory by tossing clean pages or swapping
2553 * or writing dirty pages.
2555 * Display to user space in the per-cpuset read-only file
2556 * "memory_pressure". Value displayed is an integer
2557 * representing the recent rate of entry into the synchronous
2558 * (direct) page reclaim by any task attached to the cpuset.
2561 void __cpuset_memory_pressure_bump(void)
2564 fmeter_markevent(&task_cs(current)->fmeter);
2565 task_unlock(current);
2568 #ifdef CONFIG_PROC_PID_CPUSET
2570 * proc_cpuset_show()
2571 * - Print tasks cpuset path into seq_file.
2572 * - Used for /proc/<pid>/cpuset.
2573 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2574 * doesn't really matter if tsk->cpuset changes after we read it,
2575 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2578 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2581 struct task_struct *tsk;
2583 struct cgroup_subsys_state *css;
2587 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2593 tsk = get_pid_task(pid, PIDTYPE_PID);
2599 css = task_subsys_state(tsk, cpuset_subsys_id);
2600 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2607 put_task_struct(tsk);
2614 static int cpuset_open(struct inode *inode, struct file *file)
2616 struct pid *pid = PROC_I(inode)->pid;
2617 return single_open(file, proc_cpuset_show, pid);
2620 const struct file_operations proc_cpuset_operations = {
2621 .open = cpuset_open,
2623 .llseek = seq_lseek,
2624 .release = single_release,
2626 #endif /* CONFIG_PROC_PID_CPUSET */
2628 /* Display task mems_allowed in /proc/<pid>/status file. */
2629 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2631 seq_printf(m, "Mems_allowed:\t");
2632 seq_nodemask(m, &task->mems_allowed);
2633 seq_printf(m, "\n");
2634 seq_printf(m, "Mems_allowed_list:\t");
2635 seq_nodemask_list(m, &task->mems_allowed);
2636 seq_printf(m, "\n");